GEMMA User Manual

Xiang Zhou

May 18, 2016

Contents
1 Introduction 4
1.1 What is GEMMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 How to Cite GEMMA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.1 Univariate Linear Mixed Model . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.2 Multivariate Linear Mixed Model . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.3.3 Bayesian Sparse Linear Mixed Model . . . . . . . . . . . . . . . . . . . . . . . 6
1.3.4 Variance Component Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4 Missing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4.1 Missing Genotypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.4.2 Missing Phenotypes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Installing and Compiling GEMMA 9

3 Input File Formats 10

3.1 PLINK Binary PED File Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 BIMBAM File Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2.1 Mean Genotype File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
3.2.2 Phenotype File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.3 SNP Annotation File (optional) . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3 Relatedness Matrix File Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3.1 Original Matrix Format . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.3.2 Eigen Value and Eigen Vector Format . . . . . . . . . . . . . . . . . . . . . . 14
3.4 Covariates File Format (optional) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
3.5 Beta/Z File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.6 Category File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.7 LD Score File . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

1
4 Running GEMMA 17
4.1 A Small GWAS Example Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.2 SNP filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.3 Association Tests with a Linear Model . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.3.1 Basic Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.3.2 Detailed Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.3.3 Output Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.4 Estimate Relatedness Matrix from Genotypes . . . . . . . . . . . . . . . . . . . . . . 19
4.4.1 Basic Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
4.4.2 Detailed Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.4.3 Output Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.5 Perform Eigen-Decomposition of the Relatedness Matrix . . . . . . . . . . . . . . . . 20
4.5.1 Basic Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
4.5.2 Detailed Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.5.3 Output Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.6 Association Tests with Univariate Linear Mixed Models . . . . . . . . . . . . . . . . 21
4.6.1 Basic Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
4.6.2 Detailed Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.6.3 Output Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
4.7 Association Tests with Multivariate Linear Mixed Models . . . . . . . . . . . . . . . 23
4.7.1 Basic Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.7.2 Detailed Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.7.3 Output Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23
4.8 Fit a Bayesian Sparse Linear Mixed Model . . . . . . . . . . . . . . . . . . . . . . . . 24
4.8.1 Basic Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.8.2 Detailed Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
4.8.3 Output Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
4.9 Predict Phenotypes Using Output from BSLMM . . . . . . . . . . . . . . . . . . . . 26
4.9.1 Basic Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.9.2 Detailed Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
4.9.3 Output Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.10 Variance Component Estimation with Relatedness Matrices . . . . . . . . . . . . . . 27
4.10.1 Basic Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.10.2 Detailed Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.10.3 Output Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
4.11 Variance Component Estimation with Summary Statistics . . . . . . . . . . . . . . . 28
4.11.1 Basic Usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2
 4.11.2 Detailed Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.11.3 Output Files . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

5 Questions and Answers 31

6 Options 32

3
1 Introduction
1.1 What is GEMMA
GEMMA is the software implementing the Genome-wide Efficient Mixed Model Association al-
gorithm [7] for a standard linear mixed model and some of its close relatives for genome-wide
association studies (GWAS). It fits a univariate linear mixed model (LMM) for marker association
tests with a single phenotype to account for population stratification and sample structure, and
for estimating the proportion of variance in phenotypes explained (PVE) by typed genotypes (i.e.
”chip heritability”) [7]. It fits a multivariate linear mixed model (mvLMM) for testing marker as-
sociations with multiple phenotypes simultaneously while controlling for population stratification,
and for estimating genetic correlations among complex phenotypes [8]. It fits a Bayesian sparse
linear mixed model (BSLMM) using Markov chain Monte Carlo (MCMC) for estimating PVE by
typed genotypes, predicting phenotypes, and identifying associated markers by jointly modeling all
markers while controlling for population structure [6]. It fits HE, REML and MQS for variance com-
ponent estimation using either individual-level data or summary statistics [5]. It is computationally
efficient for large scale GWAS and uses freely available open-source numerical libraries.

1.2 How to Cite GEMMA

• Software tool and univariate linear mixed models
Xiang Zhou and Matthew Stephens (2012). Genome-wide efficient mixed-model analysis for
association studies. Nature Genetics. 44: 821-824.

• Multivariate linear mixed models

Xiang Zhou and Matthew Stephens (2014). Efficient multivariate linear mixed model algo-
rithms for genome-wide association studies. Nature Methods. 11: 407-409.

• Bayesian sparse linear mixed models

Xiang Zhou, Peter Carbonetto and Matthew Stephens (2013). Polygenic modeling with
Bayesian sparse linear mixed models. PLoS Genetics. 9(2): e1003264.

• Variance component estimation with individual-level or summary data

Xiang Zhou (2016). A unified framework for variance component estimation with summary
statistics in genome-wide association studies. bioRxiv. 042846.

4
1.3 Models
1.3.1 Univariate Linear Mixed Model

GEMMA can fit a univariate linear mixed model in the following form:

y = Wα + xβ + u + ; u ∼ MVNn (0, λτ −1 K),  ∼ MVNn (0, τ −1 In ),

where y is an n-vector of quantitative traits (or binary disease labels) for n individuals; W =
(w1 , · · · , wc ) is an n × c matrix of covariates (fixed effects) including a column of 1s; α is a c-vector
of the corresponding coefficients including the intercept; x is an n-vector of marker genotypes; β is
the effect size of the marker; u is an n-vector of random effects;  is an n-vector of errors; τ −1 is the
variance of the residual errors; λ is the ratio between the two variance components; K is a known
n × n relatedness matrix and In is an n × n identity matrix. MVNn denotes the n-dimensional
multivariate normal distribution.
GEMMA tests the alternative hypothesis H1 : β 6= 0 against the null hypothesis H0 : β = 0 for
each SNP in turn, using one of the three commonly used test statistics (Wald, likelihood ratio or
score). GEMMA obtains either the maximum likelihood estimate (MLE) or the restricted maximum
likelihood estimate (REML) of λ and β, and outputs the corresponding p value.
In addition, GEMMA estimates the PVE by typed genotypes or “chip heritability”.

1.3.2 Multivariate Linear Mixed Model

GEMMA can fit a multivariate linear mixed model in the following form:

Y = WA + xβ T + U + E; G ∼ MNn×d (0, K, Vg ), E ∼ MNn×d (0, In×n , Ve ),

where Y is an n by d matrix of d phenotypes for n individuals; W = (w1 , · · · , wc ) is an n×c matrix

of covariates (fixed effects) including a column of 1s; A is a c by d matrix of the corresponding
coefficients including the intercept; x is an n-vector of marker genotypes; β is a d vector of marker
effect sizes for the d phenotypes; U is an n by d matrix of random effects; E is an n by d matrix
of errors; K is a known n by n relatedness matrix, In×n is a n by n identity matrix, Vg is a d by d
symmetric matrix of genetic variance component, Ve is a d by d symmetric matrix of environmental
variance component and MNn×d (0, V1 , V2 ) denotes the n×d matrix normal distribution with mean
0, row covariance matrix V1 (n by n), and column covariance matrix V2 (d by d).
GEMMA performs tests comparing the null hypothesis that the marker effect sizes for all
phenotypes are zero, H0 : β = 0, where 0 is a d-vector of zeros, against the general alternative
H1 : β 6= 0. For each SNP in turn, GEMMA obtains either the maximum likelihood estimate
(MLE) or the restricted maximum likelihood estimate (REML) of Vg and Ve , and outputs the
corresponding p value.

5
 In addition, GEMMA estimates the genetic and environmental correlations among phenotypes.

1.3.3 Bayesian Sparse Linear Mixed Model

GEMMA can fit a Bayesian sparse linear mixed model in the following form as well as a corre-
sponding probit counterpart:

y = 1n µ+Xβ+u+; βi ∼ πN(0, σa2 τ −1 )+(1−π)δ0 , u ∼ MVNn (0, σb2 τ −1 K),  ∼ MVNn (0, τ −1 In ),

where 1n is an n-vector of 1s, µ is a scalar representing the phenotype mean, X is an n × p matrix

of genotypes measured on n individuals at p genetic markers, β is the corresponding p-vector of the
genetic marker effects, and other parameters are the same as defined in the standard linear mixed
model in the previous section.
In the special case K = XXT /p (default in GEMMA), the SNP effect sizes can be decomposed
into two parts: α that captures the small effects that all SNPs have, and β that captures the
additional effects of some large effect SNPs. In this case, u = Xα can be viewed as the combined
effect of all small effects, and the total effect size for a given SNP is αi + βi .
There are two important hyper-parameters in the model: PVE, being the proportion of variance
in phenotypes explained by the sparse effects (Xβ) and random effects terms (u) together, and
PGE, being the proportion of genetic variance explained by the sparse effects terms (Xβ). These
two parameters are defined as follows:

V(Xβ + u)
PVE(β, u, τ ) := ,
V(Xβ + u) + τ −1
V(Xβ)
PGE(β, u) := ,
V(Xβ + u)

where
n
1X
V(x) := (xi − x)2 .
n
i=1

GEMMA uses MCMC to estimate β, u and all other hyper-parameters including PVE, PGE
and π.

6
1.3.4 Variance Component Models

GEMMA can be used to estimate variance components from a multiple-component linear mixed
model in the following form:

k
X
y= Xi β i + ; βil ∼ N(0, σi2 /pi ),  ∼ MVNn (0, σe2 In ),
i=1

which is equivalent to

k
X
y= ui + ; ui ∼ MVNn (0, σi2 Ki ),  ∼ MVNn (0, σe2 In ),
i=1

where genetic markers are classified into k non-overlapping categories; Xi is an n × pi matrix of

genotypes measured on n individuals at pi genetic markers in i’th category; β i is the corresponding
pi -vector of the genetic marker effects, where each element follows a normal distribution with
variance σi2 /pi ; ui is the combined genetic effects from i’th category; Ki = Xi XTi /pi is the category
specific genetic relatedness matrix; and other parameters are the same as defined in the standard
linear mixed model in the previous section.
GEMMA estimates the variance components σi2 . When individual-level data are available,
GEMMA uses the HE regression method or the REML average information (AI) algorithm for
estimation. When summary-level data are available, GEMMA uses MQS (MINQUE for Summary
Statistics) for estimation.

1.4 Missing Data

1.4.1 Missing Genotypes

As mentioned before [7], the tricks used in the GEMMA algorithm rely on having complete or
imputed genotype data at each SNP. That is, GEMMA requires the user to impute all missing
genotypes before association testing. This imputation step is arguably preferable than simply
dropping individuals with missing genotypes, since it can improve power to detect associations [1].
Therefore, for fitting both LMM or BSLMM, missing genotypes are recommended to be imputed
first. Otherwise, any SNPs with missingness above a certain threshold (default 5%) will not be
analyzed, and missing genotypes for SNPs that do not pass this threshold will be simply replaced
with the estimated mean genotype of that SNP. For predictions, though, all SNPs will be used
regardless of their missingness. Missing genotypes in the test set will be replaced by the mean
genotype in the training set.

7
1.4.2 Missing Phenotypes

Individuals with missing phenotypes will not be included in the LMM or BSLMM analysis. How-
ever, all individuals will be used for calculating the relatedness matrix, so that the resulting relat-
edness matrix is still an n × n matrix regardless of how many individuals have missing phenotypes.
In addition, predicted values will be calculated for individuals with missing values, based on indi-
viduals with non-missing values.
For relatedness matrix calculation, because missingness and minor allele frequency for a given
SNP are calculated based on analyzed individuals (i.e. individuals with no missing phenotypes and
no missing covariates), if all individuals have missing phenotypes, then no SNP and no individuals
will be included in the analysis and the estimated relatedness matrix will be full of “nan”s.

8
2 Installing and Compiling GEMMA
If you have downloaded a binary executable, no installation is necessary. In some cases, you may
need to use “chmod a+x gemma” before using the binary executable. In addition, notice that
the end-of-line coding in Windows (DOS) is different from that in Linux, and so you may have
to convert input files using the utility dos2unix or unix2dos in order to use them in a different
platform.
The binary executable of GEMMA works well for a reasonably large number of individuals
(say, for example, the “-eigen ” option works for at least 45,000 individuals). Due to the outdated
computation environment the software was compiled on, however, for larger sample size and for
improved computation efficiency, it is recommended to compile GEMMA on user’s own modern
computer system.
If you want to compile GEMMA by yourself, you will need to download the source code, and
you will need a standard C/C++ compiler such as GNU gcc, as well as the GSL and LAPACK
libraries. You will need to change the library paths in the Makefile accordingly. A sample Makefile
is provided along with the source code. For details on installing GSL library, please refer to
http://www.gnu.org/s/gsl/. For details on installing LAPACK library, please refer to http:
//www.netlib.org/lapack/.
If you are interested in fitting BSLMM for a large scale GWAS data set but have limited memory
to store the entire genotype matrix, you could compile GEMMA in float precision. A float precision
binary executable, named “gemmaf”, is available inside the “bin” folder in the source code. To
compile a float precision binary by yourself, you can first run “d2f.sh” script inside the “src” folder,
and then enable “FORCE FLOAT” option in the Makefile. The float version could save about half
of the memory without appreciable loss of accuracy.

9
3 Input File Formats
GEMMA requires four main input files containing genotypes, phenotypes, relatedness matrix and
(optionally) covariates. Genotype and phenotype files can be in two formats, either both in the
PLINK binary ped format or both in the BIMBAM format. Mixing genotype and phenotype files
from the two formats (for example, using PLINK files for genotypes and using BIMBAM files for
phenotypes) will result in unwanted errors. BIMBAM format is particularly useful for imputed
genotypes, as PLINK codes genotypes using 0/1/2, while BIMBAM can accommodate any real
values between 0 and 2 (and any real values if paired with “-notsnp” option). In addition, to
estimate variance components using summary statistics, GEMMA requires two other input files:
one contains marginal z-scores and the other contains SNP category.
Notice that the BIMBAM mean genotype file, the relatedness matrix file, the marginal z-score
file and the category file can be provided in compressed gzip format, while other files should be
provided in uncompressed format.

3.1 PLINK Binary PED File Format

GEMMA recognizes the PLINK binary ped file format (http://pngu.mgh.harvard.edu/~purcell/
plink/) [3] for both genotypes and phenotypes. This format requires three files: *.bed, *.bim and
*.fam, all with the same prefix. The *.bed file should be in the default SNP-major mode (beginning
with three bytes). One can use the PLINK software to generate binary ped files from standard ped
files using the following command:

plink --file [file_prefix] --make-bed --out [bedfile_prefix]

For the *.fam file, GEMMA only reads the second column (individual id) and the sixth column
(phenotype). One can specify a different column as the phenotype column by using “-n [num]”,
where ”-n 1” uses the original sixth column as phenotypes, and “-n 2” uses the seventh column,
and so on and so forth.
GEMMA codes alleles as 0/1 according to the plink webpage on binary plink format (http://
pngu.mgh.harvard.edu/~purcell/plink/binary.shtml). Specifically, the column 5 of the *.bim
file is the minor allele and is coded as 1, while the column 6 of the *.bim file is the major allele
and is coded as 0. The minor allele in column 5 is therefore the effect allele (notice that GEMMA
version 0.92 and before treats the major allele as the effect allele).
GEMMA will read the phenotypes as provided and will recognize either “-9” or “NA” as missing
phenotypes. If the phenotypes in the *.fam file are disease status, one is recommended to label
controls as 0 and cases as 1, as the results will have better interpretation. For example, the
predicted values from a linear BSLMM can be directly interpreted as the probability of being a
case. In addition, the probit BSLMM will only recognize 0/1 as control/case labels.

10
 For prediction problems, one is recommended to list all individuals in the file, but label those
individuals in the test set as missing. This will facilitate the use of the prediction function imple-
mented in GEMMA.

3.2 BIMBAM File Format

GEMMA also recognizes BIMBAM file format (http://stephenslab.uchicago.edu/software.
html) [1], which is particularly useful for imputed genotypes as well as for general covariates other
than SNPs. BIMBAM format consists of three files, a mean genotype file, a phenotype file, and an
optional SNP annotation file. We explain these files in detail below.

3.2.1 Mean Genotype File

This file contains genotype information. The first column is SNP id, the second and third columns
are allele types with minor allele first, and the remaining columns are the posterior/imputed mean
genotypes of different individuals numbered between 0 and 2. An example mean genotype file with
two SNPs and three individuals is as follows:

rs1, A, T, 0.02, 0.80, 1.50

rs2, G, C, 0.98, 0.04, 1.00

GEMMA codes alleles exactly as provided in the mean genotype file, and ignores the allele types
in the second and third columns. Therefore, the minor allele is the effect allele only if one codes
minor allele as 1 and major allele as 0.
One can use the following bash command (in one line) to generate BIMBAM mean genotype file
from IMPUTE genotype files (http://www.stats.ox.ac.uk/~marchini/software/gwas/file_
format.html)[2]:

cat [impute filename] | awk -v s=[number of samples/individuals]

’{ printf $2 "," $4 "," $5; for(i=1; i<=s; i++) printf "," $(i*3+3)*2+$(i*3+4); printf "\n" }’
> [bimbam filename]

Notice that one may need to manually input the two quote symbols ’ . Depending on the terminal,
a direct copy/paste of the above line may result in “-bash: syntax error near unexpected token ‘(’
” errors.
Finally, the mean genotype file can accommodate values other than SNP genotypes. One can
use the “-notsnp” option to disable the minor allele frequency cutoff and to use any numerical
values as covariates.

11
3.2.2 Phenotype File

This file contains phenotype information. Each line is a number indicating the phenotype value for
each individual in turn, in the same order as in the mean genotype file. Notice that only numeric
values are allowed and characters will not be recognized by the software. Missing phenotype
information is denoted as NA. The number of rows should be equal to the number of individuals
in the mean genotype file. An example phenotype file with five individuals and one phenotype is
as follows:

1.2
NA
2.7
-0.2
3.3

One can include multiple phenotypes as multiple columns in the phenotype file, and specify a
different column for association tests by using “-n [num]”, where “-n 1” uses the original first
column as phenotypes, and “-n 2” uses the second column, and so on and so forth. An example
phenotype file with five individuals and three phenotypes is as follows:

1.2 -0.3 -1.5

NA 1.5 0.3
2.7 1.1 NA
-0.2 -0.7 0.8
3.3 2.4 2.1

For binary traits, one is recommended to label controls as 0 and cases as 1, as the results will have
better interpretation. For example, the predicted values from a linear BSLMM can be directly
interpreted as the probability of being a case. In addition, the probit BSLMM will only recognize
0/1 as control/case labels.
For prediction problems, one is recommended to list all individuals in the file, but label those
individuals in the test set as missing. This will facilitate the use of the prediction function imple-
mented in GEMMA.

3.2.3 SNP Annotation File (optional)

This file contains SNP information. The first column is SNP id, the second column is its base-pair
position, and the third column is its chromosome number. The rows are not required to be in
the same order of the mean genotype file, but must contain all SNPs in that file. An example
annotation file with four SNPs is as follows:

12
rs1, 1200, 1
rs2, 1000, 1
rs3, 3320, 1
rs4, 5430, 1

If an annotation file is not provided, the SNP information columns in the output file for association
tests will have “-9” as missing values.

3.3 Relatedness Matrix File Format

GEMMA, as a linear mixed model software, requires a relatedness matrix file in addition to both
genotype and phenotype files. The relatedness matrix can be supplied in two different ways:
either use the original relatedness matrix, or use the eigen values and eigen vectors of the original
relatedness matrix.

3.3.1 Original Matrix Format

GEMMA takes the original relatedness matrix file in two formats. The first format is a n × n
matrix, where each row and each column corresponds to individuals in the same order as in the
*.fam file or in the mean genotype file, and ith row and jth column is a number indicating the
relatedness value between ith and jth individuals. An example relatedness matrix file with three
individuals is as follows:

0.3345 -0.0227 0.0103

-0.0227 0.3032 -0.0253
0.0103 -0.0253 0.3531

The second relatedness matrix format is a three column “id id value” format, where the first two
columns show two individual id numbers, and the third column shows the relatedness value between
these two individuals. Individual ids are not required to be in the same order as in the *.fam file,
and relatedness values not listed in the relatedness matrix file will be considered as 0. An example
relatedness matrix file with the same three individuals above is shown below:

id1 id1 0.3345

id1 id2 -0.0227
id1 id3 0.0103
id2 id2 0.3032
id2 id3 -0.0253
id3 id3 0.3531

As BIMBAM mean genotype files do not provide individual id, the second format only works with
the PLINK binary ped format. One can use “-km [num]” to choose which format to use, i.e. use

13
“-km 1” or “-km 2” to accompany PLINK binary ped format, and use “-km 1” to accompany
BIMBAM format.

3.3.2 Eigen Value and Eigen Vector Format

GEMMA can also read the relatedness matrix in its decomposed forms. To do this, one should
supply two files instead of one: one file containing the eigen values and the other file containing
the corresponding eigen vectors. The eigen value file contains one column of na elements, with
each element corresponds to an eigen value. The eigen vector file contains a na × na matrix, with
each column corresponds to an eigen vector. The eigen vector in the ith column of the eigen vector
file should correspond to the eigen value in the ith row of the eigen value file. Both files can be
generated from the original relatedness matrix file by using the “-eigen ” option in GEMMA. Notice
that na represents the number of analyzed individuals, which may be smaller than the number of
total individuals n.

3.4 Covariates File Format (optional)

One can provide a covariates file if needed for fitting LMM if necessary. GEMMA fits a linear
mixed model with an intercept term if no covariates file is provided, but does not internally provide
an intercept term if a covariates file is available. Therefore, if one has covariates other than the
intercept and wants to adjust for those covariates (W) simultaneously, one should provide GEMMA
with a covariates file containing an intercept term explicitly. The covariates file is similar to the
above BIMBAM multiple phenotype file, and must contain a column of 1’s if one wants to include
an intercept. An example covariates file with five individuals and three covariates (the first column
is the intercept) is as follows:

1 1 -1.5
1 2 0.3
1 2 0.6
1 1 -0.8
1 1 2.0

It can happen, especially in a small GWAS data set, that some of the covariates will be identical
to some of the genotypes (up to a scaling factor). This can cause problems in the optimization
algorithm and evoke GSL errors. To avoid this, one can either regress the phenotypes on the
covariates and use the residuals as new phenotypes, or use only SNPs that are not identical to any
of the covariates for the analysis. The later can be achieved, for example, by performing a standard
linear regression in the genotype data, but with covariates as phenotypes.

14
3.5 Beta/Z File
This file contains marginal z-scores from the study. The first row is a header line. The first column
is the SNP id, the second column is number of total SNPs, the third column is the marginal z-score,
the fourth and fifth columns are the SNP alleles. The SNPs are not required to be in the same
order of the other files. An example category file with four SNPs is as follows:

SNP N Z INC_ALLELE DEC_ALLELE

rs1 1200 -0.322165 A T
rs2 1000 -0.343634 G T
rs3 3320 -0.338341 A T
rs4 5430 -0.322820 T C

This file is flexible. You can use beta and se beta columns instead of marginal z-scores. You can
directly use the output *.assoc.txt file from the a linear model analysis as the input beta/z file.

3.6 Category File

This file contains SNP category information. The first row is a header line. The first column
is chromosome number (optional), the second column is base pair position (optional), the third
column is SNP id, the fourth column is its genetic distance on the chromosome (optional), and the
following columns list non-overlapping categories. A vector of indicators is provided for each SNP.
The SNPs are not required to be in the same order of the other files. An example category file with
four SNPs is as follows:

CHR BP SNP CM CODING UTR PROMOTER DHS INTRON ELSE

1 1200 rs1 0.428408 1 0 0 0 0 0
1 1000 rs2 0.743268 0 0 0 0 0 1
1 3320 rs3 0.744197 0 0 1 1 0 0
1 5430 rs4 0.766409 0 0 0 0 0 0

In the above file, rs1 belongs to a coding region; rs2 belongs does not belong to any of the first five
categories; rs3 belongs to both promoter and DHS regions but will be treated as an DHS snp in
the analysis; rs4 does not belong to any category and will be ignored in the analysis. Note that if
a SNP is labeled with more than one category, then it will be treated as the last category label.
This file is also flexible, as long as it contains the SNP id and the category information.

3.7 LD Score File

This file contains the LD scores for all SNPs. The first row is a header line. The first column is
chromosome number (optional), the second column is SNP id, the third column is base pair position

15
(optional), the fourth column is the LD score of the SNP. An example LD score file with four SNPs
is as follows:

CHR SNP BP L2
1 rs1 1200 1.004
1 rs2 1000 1.052
1 rs3 3320 0.974
1 rs4 5430 0.986

In the above file, the LD score for rs1 is 1.004 and the LD score for rs4 is 0.986.
This file is also flexible, as long as it contains the SNP id and the LD score information.

16
4 Running GEMMA
4.1 A Small GWAS Example Dataset
If you downloaded the GEMMA source code recently, you will find an “example” folder containing
a small GWAS example dataset. This data set comes from the heterogeneous stock mice data,
kindly provided by Wellcome Trust Centre for Human Genetics on the public domain http://mus.
well.ox.ac.uk/mouse/HS/, with detailed described in [4].
The data set consists of 1904 individuals from 85 families, all descended from eight inbred
progenitor strains. We selected two phenotypes from this data set: the percentage of CD8+ cells,
with measurements in 1410 individuals; mean corpuscular hemoglobin (MCH), with measurements
in 1580 individuals. A total of 1197 individuals have both phenotypes. The phenotypes were
already corrected for sex, age, body weight, season and year effects by the original study, and we
further quantile normalized the phenotypes to a standard normal distribution. In addition, we
obtained a total of 12,226 autosomal SNPs, with missing genotypes replaced by the mean genotype
of that SNP in the family. Genotype and phenotype files are in both BIMBAM and PLINK binary
formats.
For demonstration purpose, for CD8, we randomly divided the 85 families into two sets, where
each set contained roughly half of the individuals (i.e. inter-family split) as in [6]. We also created
artificial binary phenotypes from the quantitative phenotypes CD8, by assigning the half individuals
with higher quantitative values to 1 and the other half to 0, as in [6]. Therefore, the phenotype files
contain six columns of phenotypes. The first column contains the quantitative phenotypes CD8 for
all individuals. The second column contains quantitative phenotypes CD8 for individuals in the
training set. The third column contains quantitative phenotypes CD8 for individuals in the test
set. The fourth and fifth columns contain binary phenotypes CD8 for individuals in the training
set and test set, respectively. The sixth column contains the quantitative phenotypes MCH for all
individuals.
A demo.txt file inside the same folder shows detailed steps on how to use GEMMA to estimate
the relatedness matrix from genotypes, how to perform association tests using both the univariate
linear mixed model and the multivariate linear mixed model, how to fit the Bayesian sparse linear
mixed model and how to obtain predicted values using the output files from fitting BSLMM. The
output results from GEMMA for all the examples are also available inside the “result” subfolder.

4.2 SNP filters

The are a few SNP filters implemented in the software.

• Polymorphism. Non-polymorphic SNPs will not be included in the analysis.

• Missingness. By default, SNPs with missingness below 5% will not be included in the analysis.

17
 Use “-miss [num]” to change. For example, “-miss 0.1” changes the threshold to 10%.

• Minor allele frequency. By default, SNPs with minor allele frequency below 1% will not be
included in the analysis. Use “-maf [num]” to change. For example, “-maf 0.05” changes the
threshold to 5%.

• Correlation with any covariate. By default, SNPs with r2 correlation with any of the covariates
above 0.9999 will not be included in the analysis. Use “-r2 [num]” to change. For example,
“-r2 0.999999” changes the threshold to 0.999999.

• Hardy-Weinberg equilibrium. Use “-hwe [num]” to specify. For example, “-hwe 0.001” will
filter out SNPs with Hardy-Weinberg p values below 0.001.

• User-defined SNP list. Use “-snps [filename]” to specify a list of SNPs to be included in the
analysis.

Calculations of the above filtering thresholds are based on analyzed individuals (i.e. individuals
with no missing phenotypes and no missing covariates). Therefore, if all individuals have missing
phenotypes, no SNP will be analyzed and the output matrix will be full of “nan”s.

4.3 Association Tests with a Linear Model

4.3.1 Basic Usage

The basic usages for linear model association analysis with either the PLINK binary ped format or
the BIMBAM format are:

./gemma -bfile [prefix] -lm [num] -o [prefix]

./gemma -g [filename] -p [filename] -a [filename] -lm [num] -o [prefix]

where the “-lm [num]” option specifies which frequentist test to use, i.e. “-lm 1” performs Wald
test, “-lm 2” performs likelihood ratio test, “-lm 3” performs score test, and “-lm 4” performs
all the three tests; “-bfile [prefix]” specifies PLINK binary ped file prefix; “-g [filename]” specifies
BIMBAM mean genotype file name; “-p [filename]” specifies BIMBAM phenotype file name; “-a
[filename]” (optional) specifies BIMBAM SNP annotation file name; “-o [prefix]” specifies output
file prefix.
Notice that different from a linear mixed model, this analysis does not require a relatedness
matrix.

4.3.2 Detailed Information

For binary traits, one can label controls as 0 and cases as 1, and follow our previous approaches
to fit the data with a linear mixed model by treating the binary case control labels as quantitative

18
traits [7, 6]. This approach can be justified partly by recognizing the linear model as a first order
Taylor approximation to a generalized linear model, and partly by the robustness of the linear
model to model misspecification [6].

4.3.3 Output Files

There will be two output files, both inside an output folder in the current directory. The pre-
fix.log.txt file contains some detailed information about the running parameters and computation
time. In addition, prefix.log.txt contains PVE estimate and its standard error in the null linear
mixed model.
The prefix.assoc.txt contains the results. An example file with a few SNPs is shown below:

chr rs ps n_mis n_obs allele1 allele0 af beta se p_wald

1 rs3683945 3197400 0 1410 A G 0.443 -1.586575e-01 3.854542e-02 4.076703e-05
1 rs3707673 3407393 0 1410 G A 0.443 -1.563903e-01 3.855200e-02 5.252187e-05
1 rs6269442 3492195 0 1410 A G 0.365 -2.349908e-01 3.905200e-02 2.256622e-09
1 rs6336442 3580634 0 1410 A G 0.443 -1.566721e-01 3.857380e-02 5.141944e-05
1 rs13475700 4098402 0 1410 A C 0.127 2.209575e-01 5.644804e-02 9.497902e-05

The 11 columns are: chromosome numbers, snp ids, base pair positions on the chromosome,
number of missing individuals for a given snp, number of non-missing individuals for a given snp,
minor allele, major allele, allele frequency, beta estimates, standard errors for beta, and p values
from the Wald test.

4.4 Estimate Relatedness Matrix from Genotypes

4.4.1 Basic Usage

The basic usages to calculate an estimated relatedness matrix with either the PLINK binary ped
format or the BIMBAM format are:

./gemma -bfile [prefix] -gk [num] -o [prefix]

./gemma -g [filename] -p [filename] -gk [num] -o [prefix]

where the “-gk [num]” option specifies which relatedness matrix to estimate, i.e. “-gk 1” calculates
the centered relatedness matrix while “-gk 2” calculates the standardized relatedness matrix; “-bfile
[prefix]” specifies PLINK binary ped file prefix; “-g [filename]” specifies BIMBAM mean genotype
file name; “-p [filename]” specifies BIMBAM phenotype file name; “-o [prefix]” specifies output file
prefix.
Notice that the BIMBAM mean genotype file can be provided in a gzip compressed format.

19
4.4.2 Detailed Information

GEMMA provides two ways to estimate the relatedness matrix from genotypes, using either the
centered genotypes or the standardized genotypes. We denote X as the n × p matrix of genotypes,
xi as its ith column representing genotypes of ith SNP, x̄i as the sample mean and vxi as the sample
variance of ith SNP, and 1n as a n × 1 vector of 1’s. Then the two relatedness matrices GEMMA
can calculate are as follows:
p
1X
Gc = (xi − 1n x̄i )(xi − 1n x̄i )T ,
p
i=1
p
1X 1
Gs = (xi − 1n x̄i )(xi − 1n x̄i )T .
p vxi
i=1

Which of the two relatedness matrix to choose will largely depend on the underlying genetic
architecture of the given trait. Specifically, if SNPs with lower minor allele frequency tend to
have larger effects (which is inversely proportional to its genotype variance), then the standardized
genotype matrix is preferred. If the SNP effect size does not depend on its minor allele frequency,
then the centered genotype matrix is preferred. In our previous experience based on a limited
examples, we typically find the centered genotype matrix provides better control for population
structure in lower organisms, and the two matrices seem to perform similarly in humans.

4.4.3 Output Files

There will be two output files, both inside an output folder in the current directory. The pre-
fix.log.txt file contains some detailed information about the running parameters and computation
time, while the prefix.cXX.txt or prefix.sXX.txt contains a n × n matrix of estimated relatedness
matrix.

4.5 Perform Eigen-Decomposition of the Relatedness Matrix

4.5.1 Basic Usage

The basic usages to perform an eigen-decomposition of the relatedness matrix with either the
PLINK binary ped format or the BIMBAM format are:

./gemma -bfile [prefix] -k [filename] -eigen -o [prefix]

./gemma -g [filename] -p [filename] -k [filename] -eigen -o [prefix]

where the “-bfile [prefix]” specifies PLINK binary ped file prefix; “-g [filename]” specifies BIMBAM
mean genotype file name; “-p [filename]” specifies BIMBAM phenotype file name; “-k [filename]”
specifies the relatedness matrix file name; “-o [prefix]” specifies output file prefix.

20
 Notice that the BIMBAM mean genotype file and/or the relatedness matrix file can be provided
in a gzip compressed format.

4.5.2 Detailed Information

GEMMA extracts the matrix elements corresponding to the analyzed individuals (which may be
smaller than the number of total individuals), center the matrix, and then perform an eigen-
decomposition.

4.5.3 Output Files

There will be three output files, all inside an output folder in the current directory. The pre-
fix.log.txt file contains some detailed information about the running parameters and computation
time, while the prefix.eigenD.txt and prefix.eigenU.txt contain the eigen values and eigen vectors
of the estimated relatedness matrix, respectively.

4.6 Association Tests with Univariate Linear Mixed Models

4.6.1 Basic Usage

The basic usages for association analysis with either the PLINK binary ped format or the BIMBAM
format are:

./gemma -bfile [prefix] -k [filename] -lmm [num] -o [prefix]

./gemma -g [filename] -p [filename] -a [filename] -k [filename] -lmm [num] -o [prefix]

where the “-lmm [num]” option specifies which frequentist test to use, i.e. “-lmm 1” performs Wald
test, “-lmm 2” performs likelihood ratio test, “-lmm 3” performs score test, and “-lmm 4” performs
all the three tests; “-bfile [prefix]” specifies PLINK binary ped file prefix; “-g [filename]” specifies
BIMBAM mean genotype file name; “-p [filename]” specifies BIMBAM phenotype file name; “-
a [filename]” (optional) specifies BIMBAM SNP annotation file name; “-k [filename]” specifies
relatedness matrix file name; “-o [prefix]” specifies output file prefix.
To detect gene environmental interactions, you can add ”-gxe [filename]”. This gxe file contains
a column of environmental variables. In this case, for each SNP in turn, GEMMA will fit a linear
mixed model that controls both the SNP main effect and environmental main effect, while testing
for the interaction effect.
Notice that “-k [filename]” could be replaced by “-d [filename]” and “-u [filename]”, where “-
d [filename]” specifies the eigen value file and “-u [filename]” specifies the eigen vector file. The
BIMBAM mean genotype file and/or the relatedness matrix file (or the eigen vector file) can be
provided in a gzip compressed format.

21
4.6.2 Detailed Information

The algorithm calculates either REML or MLE estimate of λ in the evaluation interval from 1×10−5
(corresponding to almost pure environmental effect) to 1×105 (corresponding to almost pure genetic
effect). Although unnecessary in most cases, one can expand or reduce this evaluation interval by
specifying “-lmin” and “-lmax” (e.g. “-lmin 0.01 -lmax 100” specifies an interval from 0.01 to 100).
The log-scale evaluation interval is further divided into 10 equally spaced regions, and optimization
is carried out in each region where the first derivatives change sign. Although also unnecessary in
most cases, one can increase or decrease the number of regions by using “-region” (e.g. “-region
100” uses 100 equally spaced regions on the log-scale), which may yield more stable or faster
performance, respectively.
For binary traits, one can label controls as 0 and cases as 1, and follow our previous approaches
to fit the data with a linear mixed model by treating the binary case control labels as quantitative
traits [7, 6]. This approach can be justified partly by recognizing the linear model as a first order
Taylor approximation to a generalized linear model, and partly by the robustness of the linear
model to model misspecification [6].

4.6.3 Output Files

There will be two output files, both inside an output folder in the current directory. The pre-
fix.log.txt file contains some detailed information about the running parameters and computation
time. In addition, prefix.log.txt contains PVE estimate and its standard error in the null linear
mixed model.
The prefix.assoc.txt contains the results. An example file with a few SNPs is shown below:

chr rs ps n_miss allele1 allele0 af beta se l_remle p_wald

1 rs3683945 3197400 0 A G 0.443 -7.788665e-02 6.193502e-02 4.317993e+00 2.087616e-01
1 rs3707673 3407393 0 G A 0.443 -6.654282e-02 6.210234e-02 4.316144e+00 2.841271e-01
1 rs6269442 3492195 0 A G 0.365 -5.344241e-02 5.377464e-02 4.323611e+00 3.204804e-01
1 rs6336442 3580634 0 A G 0.443 -6.770154e-02 6.209267e-02 4.315713e+00 2.757541e-01
1 rs13475700 4098402 0 A C 0.127 -5.659089e-02 7.175374e-02 4.340145e+00 4.304306e-01

The 11 columns are: chromosome numbers, snp ids, base pair positions on the chromosome,
number of missing values for a given snp, minor allele, major allele, allele frequency, beta estimates,
standard errors for beta, remle estimates for lambda, and p values from Wald test.

22
4.7 Association Tests with Multivariate Linear Mixed Models
4.7.1 Basic Usage

The basic usages for association analysis with either the PLINK binary ped format or the BIMBAM
format are:

./gemma -bfile [prefix] -k [filename] -lmm [num] -n [num1] [num2] [num3] -o [prefix]
./gemma -g [filename] -p [filename] -a [filename] -k [filename] -lmm [num]
-n [num1] [num2] [num3] -o [prefix]

This is identical to the above univariate linear mixed model association test, except that an ”-n ”
option is employed to specify which phenotypes in the phenotype file are used for association tests.
(The values after the “-n ” option should be separated by a space.)
To detect gene environmental interactions, you can add ”-gxe [filename]”. This gxe file contains
a column of environmental variables. In this case, for each SNP in turn, GEMMA will fit a linear
mixed model that controls both the SNP main effect and environmental main effect, while testing
for the interaction effect.
Notice that “-k [filename]” could be replaced by “-d [filename]” and “-u [filename]”, where “-
d [filename]” specifies the eigen value file and “-u [filename]” specifies the eigen vector file. The
BIMBAM mean genotype file and/or the relatedness matrix file (or the eigen vector file) can be
provided in a gzip compressed format.

4.7.2 Detailed Information

Although the number of phenotypes used for analysis can be arbitrary, it is highly recommended
to restrict the number of phenotypes to be small, say, less than ten.
In addition, when a small proportion of phenotypes are partially missing, one can impute these
missing values before association tests:

./gemma -bfile [prefix] -k [filename] -predict -n [num1] [num2] [num3] -o [prefix]

./gemma -g [filename] -p [filename] -a [filename] -k [filename] -predict
-n [num1] [num2] [num3] -o [prefix]

4.7.3 Output Files

There will be two output files, both inside an output folder in the current directory. The pre-
fix.log.txt file contains some detailed information about the running parameters and computation
time. In addition, prefix.log.txt contains genetic correlations estimates and their standard errors
in the null multivariate linear mixed model.
The prefix.assoc.txt contains the results, and is in a very similar format as the result file from
the univariate association tests. The number of columns will depend on the number of phenotypes

23
used for analysis. The first few columns are: chromosome numbers, snp ids, base pair positions
on the chromosome, number of missing values for a given snp, minor allele, major allele and allele
frequency. The last column contains p values from the association tests. The middle columns
contain beta estimates and the variance matrix for these estimates.

4.8 Fit a Bayesian Sparse Linear Mixed Model

4.8.1 Basic Usage

The basic usages for fitting a BSLMM with either the PLINK binary ped format or the BIMBAM
format are:

./gemma -bfile [prefix] -bslmm [num] -o [prefix]

./gemma -g [filename] -p [filename] -a [filename] -bslmm [num] -o [prefix]

where the “-bslmm [num]” option specifies which model to fit, i.e. “-bslmm 1” fits a standard linear
BSLMM, “-bslmm 2” fits a ridge regression/GBLUP, and “-bslmm 3” fits a probit BSLMM; “-bfile
[prefix]” specifies PLINK binary ped file prefix; “-g [filename]” specifies BIMBAM mean genotype
file name; “-p [filename]” specifies BIMBAM phenotype file name; “-a [filename]” (optional) specifies
BIMBAM SNP annotation file name; “-o [prefix]” specifies output file prefix.
Notice that the BIMBAM mean genotype file can be provided in a gzip compressed format.

4.8.2 Detailed Information

Notice that a large memory is needed to fit BSLMM (e.g. may need 20 GB for a data set with
4000 individuals and 400,000 SNPs), because the software has to store the whole genotype matrix
in the physical memory. The float version (gemmaf) can be used to save about half of the memory
requirement without noticeable loss of accuracy.
In default, GEMMA does not require the user to provide a relatedness matrix explicitly. It
internally calculates and uses the centered relatedness matrix, which has the nice interpretation
that each effect size βi follows a mixture of two normal distributions a priori. Of course, one can
choose to supply a relatedness matrix by using the “-k [filename]” option. In addition, GEMMA
does not take covariates file when fitting BSLMM. However, one can use the BIMBAM mean
genotype file to store these covariates and use “-notsnp” option to use them.
The option “-bslmm 1” fits a linear BSLMM using MCMC, “-bslmm 2” fits a ridge regres-
sion/GBLUP with standard non-MCMC method, and “-bslmm 3” fits a probit BSLMM using
MCMC. Therefore, option “-bslmm 2” is much faster than the other two options, and option “-
bslmm 1” is faster than “-bslmm 3”. For MCMC based methods, one can use “-w [num]” to choose
the number of burn-in iterations that will be discarded, and “-s [num]” to choose the number of
sampling iterations that will be saved. In addition, one can use “-smax [num]” to choose the num-
ber of maximum SNPs to include in the model (i.e. SNPs that have addition effects), which may

24
also be needed for the probit BSLMM because of its heavier computationAL burden. It is up to
the users to decide these values for their own data sets, in order to balance computation time and
computation accuracy.
For binary traits, one can label controls as 0 and cases as 1, and follow our previous approach
to fit the data with a linear BSLMM by treating the binary case control labels as quantitative
traits [6]. This approach can be justified by recognizing the linear model as a first order Taylor
approximation to a generalized linear model. One can of course choose to fit a probit BSLMM, but
in our limited experience, we do not find appreciable prediction accuracy gains in using the probit
BSLMM over the linear BSLMM for binary traits (see a briefly discussion in [6]). This of course
could be different for a different data set.
The genotypes, phenotypes (except for the probit BSLMM), as well as the relatedness matrix
will be centered when fitting the models. The estimated values in the output files are thus for these
centered values. Therefore, proper prediction will require genotype means and phenotype means
from the individuals in the training set, and one should always use the same phenotype file (and
the same phenotype column) and the same genotype file, with individuals in the test set labeled as
missing, to fit the BSLMM and to obtain predicted values described in the next section.

4.8.3 Output Files

There will be five output files, all inside an output folder in the current directory. The prefix.log.txt
file contains some detailed information about the running parameters and computation time. In
addition, prefix.log.txt contains PVE estimate and its standard error in the null linear mixed model
(not the BSLMM).
The prefix.hyp.txt contains the posterior samples for the hyper-parameters (h, PVE, ρ, PGE,
π and |γ|), for every 10th iteration. An example file with a few SNPs is shown below:

h pve rho pge pi n_gamma

4.777635e-01 5.829042e-01 4.181280e-01 4.327976e-01 2.106763e-03 25
5.278073e-01 5.667885e-01 3.339020e-01 4.411859e-01 2.084355e-03 26
5.278073e-01 5.667885e-01 3.339020e-01 4.411859e-01 2.084355e-03 26
6.361674e-01 6.461678e-01 3.130355e-01 3.659850e-01 2.188401e-03 25
5.479237e-01 6.228036e-01 3.231856e-01 4.326231e-01 2.164183e-03 27

The prefix.param.txt contains the posterior mean estimates for the effect size parameters (α,
β|γ == 1 and γ). An example file with a few SNPs is shown below:

chr rs ps n_miss alpha beta gamma

1 rs3683945 3197400 0 -7.314495e-05 0.000000e+00 0.000000e+00
1 rs3707673 3407393 0 -7.314495e-05 0.000000e+00 0.000000e+00
1 rs6269442 3492195 0 -3.412974e-04 0.000000e+00 0.000000e+00

25
1 rs6336442 3580634 0 -8.051198e-05 0.000000e+00 0.000000e+00
1 rs13475700 4098402 0 -1.200246e-03 0.000000e+00 0.000000e+00

Notice that the beta column contains the posterior mean estimate for βi |γi == 1 rather than
βi . Therefore, the effect size estimate for the additional effect is βi γi , and in the special case
K = XXT /p, the total effect size estimate is αi + βi γi .
The prefix.bv.txt contains a column of breeding value estimates û. Individuals with missing
phenotypes will have values of “NA”.
The prefix.gamma.txt contains the posterior samples for the gamma, for every 10th iteration.
Each row lists the SNPs included in the model in that iteration, and those SNPs are represented
by their row numbers (+1) in the prefix.param.txt file.

4.9 Predict Phenotypes Using Output from BSLMM

4.9.1 Basic Usage

The basic usages for association analysis with either the PLINK binary ped format or the BIMBAM
format are:

./gemma -bfile [prefix] -epm [filename] -emu [filename] -ebv [filename] -k [filename]
-predict [num] -o [prefix]
./gemma -g [filename] -p [filename] -epm [filename] -emu [filename] -ebv [filename]
-k [filename] -predict [num] -o [prefix]

where the “-predict [num]” option specifies where the predicted values need additional transforma-
tion with the normal cumulative distribution function (CDF), i.e. “-predict 1” obtains predicted
values, “-predict 2” obtains predicted values and then transform them using the normal CDF to
probability scale; “-bfile [prefix]” specifies PLINK binary ped file prefix; “-g [filename]” specifies
BIMBAM mean genotype file name; “-p [filename]” specifies BIMBAM phenotype file name; “-epm
[filename]” specifies the output estimated parameter file (i.e. prefix.param.txt file from BSLMM);
“-emu [filename]” specifies the output log file which contains the estimated mean (i.e. prefix.log.txt
file from BSLMM); “-ebv [filename]” specifies the output estimated breeding value file (i.e. pre-
fix.bv.txt file from BSLMM); “-k [filename]” specifies relatedness matrix file name; “-o [prefix]”
specifies output file prefix.

4.9.2 Detailed Information

GEMMA will obtain predicted values for individuals with missing phenotype, and this process will
require genotype means and phenotype means from the individuals in the training set. Therefore,
use the same phenotype file (and the same phenotype column) and the same genotype file, as used
in fitting BSLMM.

26
 There are two ways to obtain predicted values: use û and β̂, or use α̂ and β̂. We note that
û and α̂ are estimated in slightly different ways, and so even in the special case K = XXT /p, û
may not equal to Xα̂. However, in this special case, these two approaches typically give similar
results based on our previous experience. Therefore, if one used the default matrix in fitting the
BSLMM, then it may not be necessary to supply “-ebv [filename]” and “-k [filename]” options, and
GEMMA can use only the estimated parameter file and log file to obtain predicted values by the
second approach. But of course, one can choose to use the first approach which is more formal, and
when do so, one needs to calculate the centered matrix based on the same phenotype column used
in BSLMM (i.e. to use only SNPs that were used in the fitting). On the other hand, if one did
not use the default matrix in fitting the BSLMM, then one needs to supply the same relatedness
matrix here again.
The option “-predict 2” should only be used when a probit BSLMM was used to fit the data.
In particular, for binary phenotypes, if one fitted the linear BSLMM then one should use the
option“-predict 1”, and use option “-predict 2” only if one fitted the data with the probit BSLMM.
Here, unlike in previous sections, all SNPs that have estimated effect sizes will be used to
obtain predicted values, regardless of their minor allele frequency and missingness. SNPs with
missing values will be imputed by the mean genotype of that SNP in the training data set.

4.9.3 Output Files

There will be two output files, both inside an output folder in the current directory. The pre-
fix.log.txt file contains some detailed information about the running parameters and computation
time, while the prefix.prdt.txt contains a column of predicted values for all individuals. In par-
ticular, individuals with missing phenotypes will have predicted values, while individuals with
non-missing phenotypes will have “NA”s.

4.10 Variance Component Estimation with Relatedness Matrices

4.10.1 Basic Usage

The basic usages for variance component estimation with relatedness matrices are:

./gemma -p [filename] -k [filename] -n [num] -vc [num] -o [prefix]

./gemma -p [filename] -mk [filename] -n [num] -vc [num] -o [prefix]

where the “-vc [num]” option specifies which estimation to use, in particular, ”-vc 1” (default)
uses HE regression and ”-vc 2” uses REML AI algorithm; “-p [filename]” specifies phenotype file
name; “-n [num]” (default 1) specifies which column of phenotype to use (e.g. one can use ”-n 6”
for a fam file); “-k [filename]” specifies relatedness matrix file name; “-mk [filename]” specifies the
multiple relatedness matrix file name; the multiple relatedness matrix file is a text file where each
row contains the full path to the relatedness matrices; “-o [prefix]” specifies output file prefix.

27
 The relatedness matrix file can be provided in a gzip compressed format.

4.10.2 Detailed Information

By default, the variance component estimates from the REML AI algorithm are constrained to be
positive. To allow for unbiased estimates, one can use ”-noconstrain” to pair with ”-vc 2”. The
estimates from the HE regression are not constrained.
For binary traits, one can label controls as 0 and cases as 1, and follow our previous approaches
to fit the data with a linear mixed model by treating the binary case control labels as quantitative
traits [6]. A scaling factor can be used to transform variance component estimates from the observed
scale back to liability scale [6].

4.10.3 Output Files

One output file will be generated inside an output folder in the current directory. This prefix.log.txt
file contains detailed information about the running parameters, computation time, as well as the
variance component/PVE estimates and their standard errors.

4.11 Variance Component Estimation with Summary Statistics

4.11.1 Basic Usage

This analysis option requires marginal z-scores from the study and individual-level genotypes from
a random subset of the study (or a separate reference panel). The marginal z-scores are provided
in a beta file while the genotypes can be provided either in the PLINK binary ped format or the
BIMBAM format. The basic usages for variance component estimation with summary statistics
are:
./gemma -beta [filename] -bfile [prefix] -vc 1 -o [prefix]
./gemma -beta [filename] -g [filename] -p [filename] -a [filename] -vc 1 -o [prefix]
where the “-vc 1” option specifies to use MQS-HEW; “-beta [filename]” specifies beta file name;
“-bfile [prefix]” specifies PLINK binary ped file prefix; “-g [filename]” specifies BIMBAM mean
genotype file name; “-p [filename]” specifies BIMBAM phenotype file name; “-a [filename]” (op-
tional) specifies BIMBAM SNP annotation file name; “-o [prefix]” specifies output file prefix. Note
that the phenotypes in the phenotype file are not used in this analysis and are only for selecting
individuals. The use of phenotypes is different from the CI1 method detailed in the next section.
To fit a multiple variance component model, you will need to add ”-cat [filename]” to provide
the SNP category file that classifies SNPs into different non-overlapping categories.
The beta file and genotype file can be provided in a gzip compressed format. In addition, to fit
MQS-LDW, you will need to add ”-wcat [filename]” together with ”-vc 2”. The ”-wcat [filename]”
option specifies the LD score file, which can be provided in a gzip compressed format.

28
4.11.2 Detailed Information

MQS-LDW uses an iterative procedure to update the variance components. It will first compute the
MQS-HEW estimates and then use these estimates to update and obtain the MQS-LDW estimates.
Therefore, there will be two outputs in the terminal, but only the final results are saved in the output
file.
By default, the standard errors for the variance component estimates are computed with the
approximate block-wise jackknife method. The jackknife method works well for unrelated individu-
als and quantitative traits. If you are interested in using the aymptotic method that is validated in
all scenarios, you need to provide genotypes and phenotypes from the study, as well as the output
files from the previous MQS run. The basic usages for using the asymptotic form to compute the
confidence intervals are

./gemma -beta [filename] -bfile [prefix] -ref [prefix] -pve [num] -ci 1 -o [prefix]
./gemma -beta [filename] -g [filename] -p [filename] -ref [prefix] -pve [num] -ci 1 -o [prefix]

In the above usages, “-ref [prefix]” specifies the prefix of the output file (including full path) from
the previous MQS fit (e.g. q and S estimates from the reference genotype files); and “-pve [num]”
specifies the pve estimates from the previous MQS fit. PLINK format files can be replaced with
BIMBAM mean genotype files. In addition, to fit MQS-LDW, one can add ”-wcat [filename]”
together with ”-ci 2”. The ”-wcat [filename]” option specifies the LD score file, which can be
provided in a gzip compressed format.
The asympototic method requires additional summary statistics besides marginal z-scores [5].
In the current implementation of the asymptotic form, we use individual level data from the study
to compute these extra summary statistics internally. Thus, at this stage, the asymptotic form
requires the genotype and phenotype files for all individuals from the study. In the near future,
we will output these extra summary statistics to facilitate consortium studies and meta-analysis.
We are currently working with some consortium studies to figure out the best way to output these
values and to make the asymptotic method easier to use.
For binary traits, one can label controls as 0 and cases as 1, and follow our previous approaches
to fit the data with a linear mixed model by treating the binary case control labels as quantitative
traits [6]. A scaling factor can be used to transform variance component estimates from the observed
scale back to liability scale [6].

4.11.3 Output Files

Five output files will be generated inside an output folder in the current directory. The prefix.log.txt
file contains detailed information about the running parameters, computation time, as well as the
variance component/PVE estimates and their standard errors. This is the main file of interest.
The prefix.S.txt file contains S estimates and their estimated errors. The prefix.q.txt file contains q

29
estimates. The prefix.Vq.txt file contains the standard errors for q. The prefix.size.txt file contains
the number of SNPs in each category and the number of individuals in the study.

30
5 Questions and Answers
1. Q: I want to perform a cross validation with my data, and want to fit BSLMM in the training
data set and obtain predicted values for individuals in the test data set. How should I prepare
the phenotype file?
A: One should always use the same phenotype and genotype files for both fitting BSLMM
and obtaining predicted values. Therefore, one should combine individuals in the training set
and test set into a single phenotype and genotype file before running GEMMA. Specifically,
in the phenotype file, one should label individuals in the training set with the true phenotype
values, and label individuals in the test set as missing (e.g. “NA”). Then, one can fit BSLMM
with the files as BSLMM only uses individuals with non-missing phenotypes (i.e. individuals
in the training set). Afterwards, one can obtain predicted values using the “-predict” option
on the same files, and the predicted values will be obtained only for individuals with missing
phenotypes (i.e. individuals in the test set). Notice that the software will still output “NA”
for individuals with non-missing phenotypes so that the number of individuals in the output
prefix.prdt.txt file will match the total sample size. Please refer to the GWAS sample data
set and some demo scripts included with the GEMMA source code for detailed examples.

31
6 Options
File I/O Related Options

• -bfile [prefix] specify input plink binary file prefix; require .fam, .bim and .bed files

• -g [filename] specify input bimbam mean genotype file name

• -p [filename] specify input bimbam phenotype file name

• -n [num] specify phenotype column in the phenotype file (default 1); or to specify which
phenotypes are used in the mvLMM analysis

• -a [filename] specify input bimbam SNPs annotation file name (optional)

• -k [filename] specify input kinship/relatedness matrix file name

• -km [num] specify input kinship/relatedness file type (default 1; valid value 1 or 2).

• -d [filename] specify input eigen value file name

• -u [filename] specify input eigen vector file name

• -c [filename] specify input covariates file name (optional); an intercept term is needed in
the covariates file

• -gxe [filename] specify input environmental covariate file name; this is only used for de-
tecting gene x environmental interactions

• -cat [filename] specify input SNP category file name; this is only used for variance compo-
nent estimation using summary statistics

• -beta [filename] specify input beta/z file name; this is only used for variance component
estimation using summary statistics

• -epm [filename] specify input estimated parameter file name

• -en [n1] [n2] [n3] [n4] specify values for the input estimated parameter file (with a header)
(default 2 5 6 7 when no -ebv -k files, and 2 0 6 7 when -ebv and -k files are supplied; n1:
rs column number; n2: estimated alpha column number (0 to ignore); n3: estimated beta
column number (0 to ignore); n4: estimated gamma column number (0 to ignore).)

• -ebv [filename] specify input estimated random effect (breeding value) file name

• -emu [filename] specify input log file name containing estimated mean

• -mu [filename] specify estimated mean value directly, instead of using -emu file

32
 • -snps [filename] specify input snps file name to only analyze a certain set of snps; contains
a column of snp ids

• -pace [num] specify terminal display update pace (default 100000).

• -o [prefix] specify output file prefix (default “result”)

SNP Quality Control Options

• -miss [num] specify missingness threshold (default 0.05)

• -maf [num] specify minor allele frequency threshold (default 0.01)

• -r2 [num] specify r-squared threshold (default 0.9999)

• -hwe [num] specify HWE test p value threshold (default 0; no test)

• -notsnp minor allele frequency cutoff is not used and so all real values can be used as
covariates

Linear Model Options

• -lm [num] specify frequentist analysis choice (default 1; valid value 1-4; 1: Wald test; 2:
likelihood ratio test; 3: score test; 4: all 1-3.)

Relatedness Matrix Calculation Options

• -gk [num] specify which type of kinship/relatedness matrix to generate (default 1; valid
value 1-2; 1: centered matrix; 2: standardized matrix.)

Eigen Decomposition Options

• -eigen specify to perform eigen decomposition of the relatedness matrix

Linear Mixed Model Options

• -lmm [num] specify frequentist analysis choice (default 1; valid value 1-4; 1: Wald test; 2:
likelihood ratio test; 3: score test; 4: all 1-3.)

• -lmin [num] specify minimal value for lambda (default 1e-5)

• -lmax [num] specify maximum value for lambda (default 1e+5)

• -region [num] specify the number of regions used to evaluate lambda (default 10)

Bayesian Sparse Linear Mixed Model Options

33
 • -bslmm [num] specify analysis choice (default 1; valid value 1-3; 1: linear BSLMM; 2: ridge
regression/GBLUP; 3: probit BSLMM.)

• -hmin [num] specify minimum value for h (default 0)

• -hmax [num] specify maximum value for h (default 1)

• -rmin [num] specify minimum value for rho (default 0)

• -rmax [num] specify maximum value for rho (default 1)

• -pmin [num] specify minimum value for log10(pi) (default log10(1/p), where p is the number
of analyzed SNPs )

• -pmax [num] specify maximum value for log10(pi) (default log10(1) )

• -smin [num] specify minimum value for —gamma— (default 0)

• -smax [num] specify maximum value for —gamma— (default 300)

• -gmean [num] specify the mean for the geometric distribution (default: 2000)

• -hscale [num] specify the step size scale for the proposal distribution of h (value between 0
and 1, default min(10/sqrt(n),1) )

• -rscale [num] specify the step size scale for the proposal distribution of rho (value between
0 and 1, default min(10/sqrt(n),1) )

• -pscale [num] specify the step size scale for the proposal distribution of log10(pi) (value
between 0 and 1, default min(5/sqrt(n),1) )

• -w [num] specify burn-in steps (default 100,000)

• -s [num] specify sampling steps (default 1,000,000)

• -rpace [num] specify recording pace, record one state in every [num] steps (default 10)

• -wpace [num] specify writing pace, write values down in every [num] recorded steps (default
1000)

• -seed [num] specify random seed (a random seed is generated by default)

• -mh [num] specify number of MH steps in each iteration (default 10; requires 0/1 phenotypes
and -bslmm 3 option)

Prediction Options

34
 • -predict [num] specify prediction options (default 1; valid value 1-2; 1: predict for individu-
als with missing phenotypes; 2: predict for individuals with missing phenotypes, and convert
the predicted values using normal CDF.)

Variance Component Estimation Options

• -vc [num] specify fitting algorithm. For individual level data (default 1; valid value 1-2; 1:
HE regression; 2: REML AI algorithm.). For summary statistics (default 1; valid value 1-2;
1: MQS-HEW; 2: MQS-LDW.)

• -ci [num] specify fitting algorithm to compute the standard errors. (default 1; valid value
1-2; 1: MQS-HEW; 2: MQS-LDW.)

35
References
[1] Yongtao Guan and Matthew Stephens. Practical issues in imputation-based association map-
ping. PLoS Genetics, 4:e1000279, 2008.

[2] Bryan N. Howie, Peter Donnelly, and Jonathan Marchini. A flexible and accurate genotype
imputation method for the next generation of genome-wide association studies. PLoS Genetics,
5:e1000529, 2009.

[3] Shaun Purcell, Benjamin Neale, Kathe Todd-Brown, Lori Thomas, Manuel A. R. Ferreira,
David Bender, Julian Maller, Pamela Sklar, Paul I. W. de Bakker, Mark J. Daly, and Pak C.
Sham. PLINK: a toolset for whole-genome association and population-based linkage analysis.
The American Journal of Human Genetics, 81:559–575, 2007.

[4] William Valdar, Leah C. Solberg, Dominique Gauguier, Stephanie Burnett, Paul Klenerman,
William O. Cookson, Martin S. Taylor, J Nicholas P. Rawlins, Richard Mott, and Jonathan
Flint. Genome-wide genetic association of complex traits in heterogeneous stock mice. Nature
Genetics, 38:879–887, 2006.

[5] Xiang Zhou. A unified framework for variance component estimation with summary statistics
in genome-wide association studies. bioRxiv, 2016.

[6] Xiang Zhou, Peter Carbonetto, and Matthew Stephens. Polygenic modelling with Bayesian
sparse linear mixed models. PLoS Genetics, 9:e1003264, 2013.

[7] Xiang Zhou and Matthew Stephens. Genome-wide efficient mixed-model analysis for association
studies. Nature Genetics, 44:821–824, 2012.

[8] Xiang Zhou and Matthew Stephens. Efficient multivariate linear mixed model algorithms for
genome-wide association studies. Nature Methods, 11:407–409, 2014.

36